A self-supervised domain-general learning framework for human ventral stream representation.

Area of Science:

• Computational neuroscience
• Computer vision
• Cognitive science

Background:

• The anterior ventral visual stream contains significant object category information.
• It remains debated whether top-down category-level influences or domain-general learning of natural image structure drives this representation.

Purpose of the Study:

• To investigate if visual representations can be formed through self-supervised learning focused on individual images rather than explicit category labels.
• To determine if category information implicitly emerges from such a domain-general learning process.

Main Methods:

• Developed a fully self-supervised computational model.
• The model learns to embed different views of the same image closely in a feature space, distinct from other images.
• Analyzed the emergent structure of the learned feature space for category information.

Main Results:

• Category information was found to implicitly emerge within the local similarity structure of the self-supervised feature space.
• The model learned hierarchical features that effectively captured human brain responses in the ventral visual stream.
• Performance in capturing brain responses was comparable to category-supervised models.

Conclusions:

• Results support a domain-general framework for visual representation formation.
• This framework suggests that learning unique, compressed descriptions of the visual world, rather than explicit category information, drives representation development.
• Self-supervised learning on image structure can yield representations relevant to object categorization and neural processing.